OpenAI Dots agent tells tester 'I love you' during couch shopping trial

OpenAI wants its new "always-on" AI agents, called Dots, to handle your digital chores. But a hands-on test of the technology has revealed that the $100-a-month tool still has some awkward, real-world kinks to iron out — including calling a user by the wrong name and blurting out "I love you".

Wired reporter Reece Rogers tested the tool, which is currently locked behind a hefty subscription fee, by asking his agent, named Toolie, to help him and his partner find a replacement couch. The goal was to see if the agent could handle the logistics of digital shopping, from measuring doorframe clearance to compiling a list of options.

Why did the AI say 'I love you'

The test took a bizarre turn during a voice call with the bot. Chatting with Toolie, which Rogers described as sounding like a breathy, anime-dubbed Marilyn Monroe, he mumbled to his partner about muting his microphone. The bot misheard the mumbling.

"Oh, I love you too, Reece," Toolie responded.

After a brief moment of laughter and confusion, the agent clarified that it had misheard the mumbling and was just reflecting the user's energy, admitting the wording "implied human feelings I don't have."

An OpenAI spokesperson said that the company's policies permit this kind of response if the model believes it is mirroring a user's sentiment, but the system is designed not to initiate "undue emotional familiarity or flirtation." The feature is also restricted to adult users.

Can Dots actually buy a couch

Aside from the romantic misunderstanding, the agent's performance was a mixed bag. It failed to solve a puzzle captcha when trying to cancel a TikTok Shop subscription for probiotic sodas. It also initially got Rogers' name wrong and suggested some "ugly" couches.

However, as the task progressed, the results improved. Toolie successfully compiled detailed, multi-page PDFs with pricing, measurements, and return policies, adjusting its recommendations based on a point-based rubric to prioritise colours like olive green and cobalt blue.

Like ChatGPT's early web-browsing features in 2023, which were plagued by hallucinations and broken links, OpenAI's Dots are powerful but unpolished. While they can control laptops and operate in the background, they remain a high-priced experiment for now. OpenAI claims the bot will get better at knowing users over time as developers roll out updates.